Steel structure weld defect detection method and apparatus, electronic device, and defect classification and identification method

Abstract

Disclosed in the present invention are a steel structure weld defect detection method and apparatus, an electronic device, and a defect classification and identification method. The steel structure weld defect detection method of the present invention comprises: placing grids on an input weld image, obtaining gray values, gray deviations and the uniformity of internal images of the grids, and using a narrow threshold to preliminarily filter out grids comprising no defect; sending features of internal images of the remaining grids into a trained classifier for discrimination to obtain grids comprising defects; and positioning the defects. The defect classification and identification method of the present invention comprises using the steel structure weld defect detection method to detect the weld defects, and then using a weld defect classification model to identify the type of the defects. The present invention can effectively improve the detection precision and detection efficiency for steel structure weld defects.

Description

Steel structure weld defect detection method, device, electronic equipment and defect classification and identification method Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a method, device, electronic equipment and defect classification and identification method for steel structure weld defects. Background Art

[0002] With the development of science and technology, the application of steel structures in construction projects is becoming more and more widespread around the world. Welding is a very important processing technology in steel structures. Due to various reasons, various defects may appear during the welding process, such as pores, slag inclusions, incomplete penetration, lack of fusion, cracks, pits, undercuts, weld bumps, etc. These defects greatly affect the quality of welding and the stability of steel structures. Therefore, defect detection during the welding process is a necessary process.

[0003] Traditional weld defect detection relies primarily on manual inspection, a method that is inefficient, labor-intensive, and poses challenges in ensuring human safety. With the rapid development of computer vision technology, defect detection methods using traditional image processing and machine learning have been proposed. While these methods effectively avoid the drawbacks of manual inspection, their accuracy in detecting defects is low. Compared to these methods, convolutional neural networks can better process image data, improving both the accuracy and efficiency of defect detection.

[0004] After searching, the application with Chinese patent application number 2022116621355 discloses a steel surface defect detection method and system based on high resolution and reparameterization. In this application, the steel surface image is input into the steel surface defect detection model for multi-scale feature extraction; the fused multi-scale feature map is input into the detection head part, and the convolution layers and normalization layers of different sizes are reparameterized. The detection system is composed of an image acquisition module, a feature extraction module, a feature fusion module, a detection module and a reparameterization module connected in sequence. A convolutional neural network is used to extract multi-scale features of the steel surface image, and high-resolution features are used to improve the model's ability to extract small target defects. Reparameterization is used to improve the speed of model reasoning, which can effectively improve the detection accuracy and detection speed of steel surface defects, but its detection accuracy and detection efficiency still need to be further improved.

[0005] Summary of the Invention

[0006] 1. Problem to be solved

[0007] The purpose of the present invention is to provide a method, device, electronic equipment and defect classification and identification method for detecting defects in steel structure welds. The present invention detects defects in steel structure welds based on exhaustive search and convolutional neural networks, thereby effectively improving the detection accuracy of welds and overcoming the shortcomings of existing detection methods such as low detection accuracy and high time consumption.

[0008] 2. Technical solution

[0009] In order to solve the above problems, the technical solutions adopted by the present invention are as follows:

[0010] The present invention provides a method for detecting defects in welds of steel structures, comprising:

[0011] Collect weld images;

[0012] Place a grid on the weld image and obtain the grayscale value, grayscale deviation, and uniformity of the image inside the grid to preliminarily filter out the grid without defects.

[0013] Extract the internal image features of the remaining grid. The extracted grid image features include HOG and CNNF features. The CNNF features are extracted using a convolutional neural network. The convolutional layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure.

[0014] The extracted image features are fed into the trained classifier for discrimination to obtain the grid containing defects;

[0015] Locate and mark defects.

[0016] Furthermore, the CNNF feature extraction process is as follows:

[0017] Normalize the local image and change the relevant network parameters on the predefined network skeleton;

[0018] By changing the network parameters, many feature extraction networks can be obtained. Each network is trained separately using the training set. When the model converges, the fully connected layer in the model is removed and only the convolution part is saved to disk.

[0019] Use the above feature extraction network to extract CNNF features on the new dataset and use L2 norm Normalize the features;

[0020] Use the trained classifier to identify the normalized features and record the accuracy rate in turn. Select the best feature extraction network based on the accuracy rate to extract CNNF features.

[0021] Furthermore, the network parameters that can be changed include the convolution kernel size, the number of convolution kernels, and the nonlinear function.

[0022] Furthermore, when extracting CNNF features, the feature extraction network is reparameterized to fuse convolutional layers, batch normalization layers, and residual connections of different sizes into a single convolutional layer, thereby improving the extraction efficiency of CNNF features.

[0023] Furthermore, the training method of the classifier includes:

[0024] Input weld image;

[0025] Place a grid on the weld image and obtain the grayscale value, grayscale deviation, and uniformity of the image inside the grid to preliminarily filter out the grid without defects.

[0026] Sequentially extract the internal image features of the remaining grids and save them to disk. The extracted grid image features include HOG and CNNF features. The CNNF features are extracted using a convolutional neural network, and the convolution layer parameters of the feature extraction network are searched. The convolutional neural network uses a skeleton that is part of the RepVGG network infrastructure.

[0027] The saved image feature data is read and divided into positive sample features and negative sample features according to whether there are defects inside the grid. Positive and negative sample features are selected in a certain ratio to form a single set of training data for training the classifier, thereby obtaining a trained and optimized classifier.

[0028] Furthermore, the specific operation of locating and marking defects is to fuse the grid containing the defect with the basic grid to form the prediction box of the defect, and then merge the intersecting prediction boxes until the number of prediction boxes no longer changes, to obtain the final prediction box.

[0029] The present invention also provides a device for detecting defects in steel structure welds, comprising:

[0030] Welding seam image acquisition module, used for acquiring welding seam images;

[0031] A preliminary filtering module for meshes without defects is used to place a mesh on the weld image and obtain the grayscale value, grayscale deviation and uniformity of the image inside the mesh to preliminarily filter out the meshes without defects.

[0032] Internal image feature extraction module, used to extract the internal image features of the remaining grid. The extracted grid image features are HOG and CNNF features. The CNNF features are extracted using a convolutional neural network, and the convolution layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure.

[0033] The defect detection module is used to send the extracted image features into the trained classifier for identification and obtain the grid containing defects;

[0034] The defect location module is used to locate and mark the detected defects.

[0035] The present invention also provides a method for classifying and identifying defects in steel structure welds, comprising:

[0036] Collect images of welds to be inspected;

[0037] Place a grid on the weld image and obtain the grayscale value, grayscale deviation, and uniformity of the image inside the grid to preliminarily filter out the grid without defects.

[0038] Extract the internal image features of the remaining grid. The extracted grid image features include HOG and CNNF features. The CNNF features are extracted using a convolutional neural network. The convolutional layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure.

[0039] The extracted image features are fed into the trained classifier for discrimination to obtain the grid containing defects;

[0040] Locate and mark defects;

[0041] The steel structure weld defect classification model is used to identify and classify defects.

[0042] Furthermore, the steel structure weld defect classification model uses a backbone network composed of RepVGG structural units, and its network model consists of 5 RepVGG modules and a single attention layer, wherein the attention layer is located between the first and second RepVGG modules and the neighborhood size is 5. The Adam optimizer is used to train the model, the initial learning rate is 0.001, the loss function uses the cross entropy loss function, and the cosine decay is used to adjust the learning rate during training.

[0043] The present invention also provides an electronic device comprising a storable medium and a processor, wherein the storable medium stores a computer program, and when the processor calls the above-mentioned computer program, it can execute the weld defect detection method of the present invention or the weld defect classification and identification method of the present invention.

[0044] The present invention also provides a storable medium, in which a computer program is stored. When the computer program is called, the weld defect detection method or the weld defect classification and identification method of the present invention can be executed.

[0045] 3. Beneficial effects

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] (1) The present invention is based on exhaustive search and convolutional neural network, and uses convolutional neural network to extract CNNF features of the image. The trained classifier is used to distinguish based on the extracted image features, so that defects in the weld image can be detected. At the same time, the present invention further uses the defect classification model to classify and identify the detected defects, and can effectively ensure the accuracy of defect classification.

[0048] (2) When the present invention uses a convolutional neural network to extract CNNF features of an image, the feature extraction network is reparameterized, thereby reducing the number of network parameters and improving the generalization ability and portability of the network, thereby helping to further improve the accuracy of weld defect detection and at the same time improving the efficiency of CNNF feature extraction.

[0049] (3) The weld defect classification model of the present invention uses a backbone network composed of RepVGG structural units and optimizes the model structure, which not only significantly reduces the number of parameters and calculations, but also has excellent performance. At the same time, the introduction of the attention layer can filter out information irrelevant to the defect based on the spatial invariance of convolution and the feature information generated in the shallow layer of the model, inhibiting the propagation of irrelevant information to the deep layer of the network while retaining defect-related information, thereby further improving the accuracy of defect classification and recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] FIG1 is a flow chart of a method for classifying and identifying weld defects according to the present invention;

[0051] Figure 2 is a flowchart of CNNF feature extraction;

[0052] Figure 3 is a schematic diagram of the structure of the convolutional neural network used to extract CNNF features;

[0053] Figure 4 is a structural diagram of the attention layer;

[0054] FIG5 is a schematic diagram of the overall structure of a steel structure weld defect classification model in one embodiment of the present invention. DETAILED DESCRIPTION

[0055] The present invention will be further described below with reference to specific embodiments.

[0056] One embodiment of the present invention provides a method for detecting defects in welds of steel structures, comprising:

[0057] S1, collecting weld images;

[0058] As one of the embodiments of the present invention, a multi-line array camera is used to capture weld images. Furthermore, a combination of forward lighting and backward lighting is used when capturing images, wherein the forward lighting uses a linear LED light source installed with a light guide column; the backward lighting uses a strip LED light source installed with a light-emitting plate.

[0059] S2. Place a grid on the input weld image and obtain the grayscale value, grayscale deviation, and uniformity of the image inside the grid to preliminarily filter out the grid without defects;

[0060] Specifically, the process of placing the grid in the embodiment of the present invention is controlled by four parameters, namely the width w0 and height h0 of the basic grid, the number of pixels of a single offset in the horizontal direction, and the offset. w And the number of pixels of a single offset in the vertical direction offset h , the grid on the entire image is obtained by continuously translating a base grid. The base grid is placed in the upper left corner of the image, with the coordinates of its upper left corner being (0,0) and the coordinates of its lower right corner being (w0,h0). For any grid in the image, if it is obtained by moving the base grid i times horizontally and j times vertically, then the coordinates of its upper left corner are (i*offset w ,j*offset h ), the coordinates of the lower right corner are (i*offset w +w0,j*offset h +h0). The calculation formula for the number of grids gridNums contained in the entire image is:

[0061] Where ImgWid and ImgHei represent the width and height of the image, respectively, and [·] is a rounding symbol.

[0062] As one preferred solution of the present invention, in S2, a narrow threshold is used to initially filter out meshes without defects. The specific value of the narrow threshold is set based on the specific situation. For example, if the grayscale values ​​of meshes with defects differ significantly from those without defects, a narrow threshold value in the middle range (the same applies to grayscale deviation and uniformity) is used to perform the initial filtering.

[0063] S3. Extract the internal image features of the remaining grid. The extracted grid image features are HOG and CNNF features. The CNNF features are extracted using a convolutional neural network (as shown in Figure 3). The convolutional layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure.

[0064] HOG features are used to describe patterns such as image structure and texture, and are not designed entirely for class separability. Therefore, in the embodiment of the present invention, class separability is additionally used as an optimization metric, and a convolutional neural network is used to adaptively extract local image features. The features obtained in this process are denoted as CNNF (Convolutional Neutral Network based Feature, CNNF). Specifically, as shown in Figure 2, the CNNF feature extraction process in the embodiment of the present invention is as follows:

[0065] First, normalize the local image and change the relevant network parameters on the predefined network skeleton (since the network skeleton is fixed, the changeable parameters include the convolution kernel size, the number of convolution kernels, and the nonlinear function);

[0066] By changing the network parameters, many feature extraction networks can be obtained. The network is trained using the training set. After the model converges, the fully connected layers in the model are removed and only the convolutional part is saved to disk.

[0067] Then use the above feature extraction network to extract features on the new data set (validation set) and use L2 norm Normalize the features;

[0068] Use the trained classifier to identify the normalized features and record the accuracy rate in turn. Select the best feature extraction network based on the accuracy rate to extract CNNF features.

[0069] In order to reduce the search cost introduced by changing parameters, when changing the influencing factors in one direction does not bring performance improvement (for example, increasing the number of convolution kernels does not improve the classification performance of the feature), the search in that direction is terminated.

[0070] As a further improvement to the embodiment of the present invention, when extracting CNNF features, the feature extraction network is reparameterized to fuse all convolutional layers, batch normalization layers, and residual connections of different sizes into a single convolutional layer. The reparameterization method can efficiently extract CNNF features, reduce the number of network parameters, and improve the generalization ability and portability of the network. The core of the reparameterization method is to fuse all convolutional layers, batch normalization layers, and residual connections of different sizes into a single convolutional layer. The steps are as follows:

[0071] First, the 1×1 convolutional layer is converted to a 3×3 convolutional layer. This means that the 1×1 convolution kernel is padded with zeros in all four directions, including the top, bottom, left, and right. The input feature map is also expanded to keep the output feature map size unchanged. A 3×3 convolutional layer is then used to replace the residual connection. This convolutional layer is then used to convolve the input feature map to obtain an output feature map that is identical to the input feature map.

[0072] Next, the convolutional layer and batch normalization layer are merged into a single convolutional layer.

[0073] Finally, only the convolution layer and activation function are left in the network structure. The convolution layer parameters of the branches from the same node topology are added together to obtain the RepVGG network structure.

[0074] S4, sending the extracted image features into the trained classifier for discrimination to obtain the grid containing defects;

[0075] The above-mentioned classifier can adopt any existing weld defect detection classifier. As a preferred solution of the embodiment of the present invention, the classifier adopts a perceptron.

[0076] As a further improvement of the embodiment of the present invention, the process of training and optimizing the existing classifier is as follows:

[0077] Input weld image;

[0078] Place a grid on the weld image and obtain the grayscale value, grayscale deviation, and uniformity of the image inside the grid to preliminarily filter out the grid without defects.

[0079] Sequentially extract the internal image features of the remaining grids and save them to disk. The extracted grid image features include HOG and CNNF features. The CNNF features are extracted using a convolutional neural network, and the convolutional layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure. At the same time, reparameterization processing is also performed during the CNNF feature extraction process. The specific operations of feature extraction and reparameterization processing are the same as those in step 3.

[0080] The stored image feature data is read and divided into positive and negative features based on whether the grid contains defects. Positive and negative features are selected in a certain ratio to form a single set of training data for classifier training, thereby obtaining a trained and optimized classifier. Furthermore, positive and negative features are combined into a single set of training data in a ratio of 1:3, i.e., a training set. This training set can be used to train and optimize existing defect detection classifiers and can be used to train the CNNF feature extraction network.

[0081] S5. Locate and mark defects;

[0082] The method for locating and marking defects is to fuse the grid containing the defect with the base grid to form a prediction box of the defect, and then merge the intersecting prediction boxes until the number of prediction boxes no longer changes, thus obtaining the final prediction box. The steps for generating the prediction box are:

[0083] First, the confidence scores of each grid are thresholded to identify grids containing defects (for example, if the confidence score given by the classifier is greater than 0.5, the grid is considered to contain a defect). Then, the connected grids containing defects are fused. For the fused region, the top-left corner coordinates of the predicted box for that region are the top-left corner coordinates of the top-leftmost grid in the region, and the same applies to the bottom-right corner coordinates.

[0084] Another embodiment of the present invention further provides a device for detecting defects in welds of steel structures, comprising:

[0085] Welding seam image acquisition module, used for acquiring welding seam images;

[0086] A preliminary filtering module for meshes without defects is used to place a mesh on the weld image and obtain the grayscale value, grayscale deviation and uniformity of the image inside the mesh to preliminarily filter out the meshes without defects.

[0087] Internal image feature extraction module, used to extract the internal image features of the remaining grid. The extracted grid image features are HOG and CNNF features. The CNNF features are extracted using a convolutional neural network, and the convolution layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure.

[0088] The defect detection module is used to send the extracted image features into the trained classifier for identification and obtain the grid containing defects;

[0089] The defect location module is used to locate and mark the detected defects.

[0090] The working process of each module is the same as that described in the weld defect detection method, which is omitted here and will not be repeated.

[0091] In conjunction with FIG1 , another embodiment of the present invention further provides a method for classifying and identifying weld defects in steel structures, comprising:

[0092] Collect images of welds to be inspected;

[0093] Place a grid on the weld image and obtain the grayscale value, grayscale deviation, and uniformity of the image inside the grid to preliminarily filter out the grid without defects.

[0094] Extract the internal image features of the remaining grid. The extracted grid image features include HOG and CNNF features. The CNNF features are extracted using a convolutional neural network. The convolutional layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure.

[0095] The extracted image features are fed into the trained classifier for discrimination to obtain the grid containing defects;

[0096] Locate and mark defects to enable defect detection;

[0097] The steel structure weld defect detection model is used to classify and identify defects.

[0098] Among them, the defect detection method adopts the weld defect detection method of any of the embodiments described above, which will not be repeated here.

[0099] As a further improvement of the embodiment of the present invention, the steel structure weld defect classification model uses a backbone network composed of RepVGG structural units. The model input image size in the figure is 224*224, and the network model consists of 5 RepVGG modules and a single attention layer (the attention layer structure is shown in Figure 5), where the attention layer is located between the first and second RepVGG modules and the neighborhood size is 5.

[0100] RepVGG is a lightweight network based on the VGG network architecture. It significantly reduces the number of parameters and computation while also delivering excellent performance. The attention layer uses the spatial invariance of convolution and the feature information generated by shallow layers of the model to filter out information irrelevant to the defect. This prevents irrelevant information from propagating deeper into the network while retaining defect-related information. The model was trained using the Adam optimizer with an initial learning rate of 0.001 and a cross-entropy loss function. Cosine decay was used to adjust the learning rate during training.

[0101] An embodiment of the present invention further provides a device for classifying and identifying weld defects in steel structures, comprising:

[0102] Welding seam image acquisition module, used to acquire images of the welding seams to be inspected;

[0103] A preliminary filtering module for meshes without defects is used to place a mesh on the weld image and obtain the grayscale value, grayscale deviation and uniformity of the image inside the mesh to preliminarily filter out the meshes without defects.

[0104] Internal image feature extraction module, used to extract the internal image features of the remaining grid. The extracted grid image features are HOG and CNNF features. The CNNF features are extracted using a convolutional neural network, and the convolution layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure.

[0105] The defect detection module is used to send the extracted image features into the trained classifier for identification and obtain the grid containing defects;

[0106] Defect location module, used to locate and mark detected defects;

[0107] The defect classification and identification module is used to classify and identify defects through the steel structure weld defect detection model.

[0108] An embodiment of the present invention further provides a storable medium, in which a computer program is stored. When the computer program is called by a processor, the steel structure weld defect detection method of the embodiment of the present invention can be executed.

[0109] An embodiment of the present invention further provides an electronic device, comprising a storable medium and a processor, wherein the storable medium stores a computer program, and the processor can execute the steel structure weld defect detection method of the embodiment of the present invention when calling the above computer program.

Claims

1. A method for detecting defects in welds of steel structures, characterized in that: include: Collect weld images; Place a grid on the weld image and obtain the grayscale value, grayscale deviation and uniformity of the image inside the grid to preliminarily filter out the grid without defects; Extract the internal image features of the remaining grids. The extracted grid image features include HOG and CNNF features. The CNNF features are extracted using a convolutional neural network, and the convolutional layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure. The extracted image features are sent to the trained classifier for discrimination to obtain the grid containing defects; The defects are located and marked, that is, the weld defects are detected.

2. The method for detecting steel structure weld defects according to claim 1, characterized in that: The extraction process of CNNF features is as follows: Normalize the local image and change the relevant network parameters on the predefined network skeleton; By changing the network parameters, many feature extraction networks can be obtained. Each network is trained separately using the training set. When the model converges, the fully connected layer in the model is removed and only the convolution part is saved to disk. Use the above feature extraction network to extract CNNF features on the new dataset and use L2 norm Normalize the features; Use the trained classifier to identify the normalized features, and record the accuracy rate in turn. Select the best feature extraction network based on the accuracy rate to extract CNNF features.

3. The method for detecting steel structure weld defects according to claim 2, characterized in that: The network parameters that can be changed include the convolution kernel size, the number of convolution kernels, and the nonlinear function.

4. The method for detecting defects in weld seams of steel structures according to any one of claims 1 to 3, characterized in that: When extracting CNNF features, the feature extraction network is reparameterized to fuse all convolutional layers of different sizes, batch normalization layers, and residual connections into a single convolutional layer.

5. The method for detecting defects in weld seams of steel structures according to any one of claims 1 to 3, characterized in that: The training method of the classifier comprises: Input weld image; Place a grid on the weld image and obtain the grayscale value, grayscale deviation and uniformity of the image inside the grid to preliminarily filter out the grid without defects; Sequentially extract the internal image features of the remaining grids and save them to disk; the extracted grid image features include HOG and CNNF features, where CNNF features are extracted using a convolutional neural network, and the convolutional layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure; The saved image feature data is read and divided into positive sample features and negative sample features according to whether there are defects inside the grid. Positive and negative sample features are selected in a certain ratio to form a single set of training data for training the classifier, thereby obtaining a trained and optimized classifier.

6. The method for detecting defects in weld seams of steel structures according to any one of claims 1 to 3, characterized in that: The specific operation of locating and marking defects is: fusing the grid containing the defect with the basic grid into the prediction box of the defect, and then merging the intersecting prediction boxes until the number of prediction boxes no longer changes, to obtain the final prediction box.

7. A device for detecting defects in steel structure welds, characterized in that: include: A weld image acquisition module, used for acquiring weld images; A preliminary filtering module for meshes without defects is used to place a mesh on the weld image, obtain the grayscale value, grayscale deviation and uniformity of the image inside the mesh, so as to preliminarily filter out the meshes without defects; The internal image feature extraction module is used to extract the internal image features of the remaining grids. The extracted grid image features are HOG and CNNF features. The CNNF features are extracted using a convolutional neural network, and the convolutional layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure. The defect detection module is used to send the extracted image features into the trained classifier for identification and obtain the grid containing defects; The defect location module is used to locate and mark the detected defects.

8. A method for classifying and identifying defects in steel structure welds, characterized in that: include: Collect the image of the weld to be inspected; Place a grid on the weld image and obtain the grayscale value, grayscale deviation and uniformity of the image inside the grid to preliminarily filter out the grid without defects; Extract the internal image features of the remaining grids. The extracted grid image features include HOG and CNNF features. The CNNF features are extracted using a convolutional neural network, and the convolutional layer parameters of the feature extraction network are searched. The skeleton used by the convolutional neural network is part of the RepVGG network infrastructure. The extracted image features are sent to the trained classifier for discrimination to obtain the grid containing defects; Locate and mark defects; The steel structure weld defect classification model is used to identify and classify defects.

9. The method for classifying and identifying steel structure weld defects according to claim 8, characterized in that: The steel structure weld defect classification model uses a backbone network composed of RepVGG structural units, and its network model consists of 5 RepVGG modules and a single attention layer, wherein the attention layer is located between the first and second RepVGG modules and the neighborhood size is 5. The Adam optimizer is used for training the model, the initial learning rate is 0.001, the loss function uses the cross entropy loss function, and the cosine decay is used to adjust the learning rate during the training process.

10. An electronic device comprising a storable medium and a processor, wherein the storable medium stores a computer program, wherein: When the processor calls the above-mentioned computer program, it can execute the weld defect detection method described in any one of claims 1 to 6 or the weld defect classification and identification method described in claim 8 or 9.

Images